chore(deps): adapt to faststream 0.7.6 and raise its floor

[lesnik512]

Dependency refresh. uv.lock is gitignored, so the baseline is a local lock; against the live index
only these moved:

package	old	new
faststream	0.7.5	0.7.6
fast-depends	3.0.8	3.0.9
sqlalchemy	2.0.53	2.0.54
ty	0.0.81	0.0.82
ruff	0.16.7	0.16.8
eof-fixer	0.4.2	0.4.3
idna	3.19	3.20

fastapi stays at 0.139.2 under the existing <0.140 cap; lifting it is untested here.

faststream 0.7.6 accounted for all of it: 5 test failures and 15 ty diagnostics.

Upstream changes and the adaptations

The ASGI try-it-out registry monkeypatch is gone. 0.7.6 replaced the private
_get_broker_registry with a TestBroker.__init_subclass__(broker=...) hook, so TestOutboxBroker
declares broker=OutboxBroker and the 29-line try/except in __init__.py is deleted. Net removal,
and one fewer reach into a private symbol.

SubscriberSpec / PublisherSpec gained a required address, and channel_labels became
abstract. Both implemented. The subscriber keeps its single comma-joined channel rather than
splitting one per queue the way the built-in brokers now do; that is a document-shape decision
rather than a dependency one, written up with a verified blast radius in #181.

__aiter__ is a plain def upstream again (declaring it async typed it as a coroutine, which
async for rejects). Shape matched; the unsupported-peek behaviour is unchanged.

dependencies narrowed from Iterable to Sequence across upstream's signatures. Followed at
the four sites in broker.py, router.py (x2), registrator.py.

publisher.mock now raises SetupError outside a test broker instead of answering with a reset
mock. That exposed a vacuous assertion: test_relay_dual_fire_guard_through_worker_loop_leaves_row_and_logs
called assert_not_called() after leaving the TestKafkaBroker context, where the mock had already
been cleared, so it could not fail. Moved inside the context, it tests something.

Four ty: ignore suppressions became redundant and are removed with their comments.

The floor was wrong

TestBroker did not accept the broker= kwarg before 0.7.6, so on 0.7.5 import faststream_outbox
now fails at import:

TypeError: TestOutboxBroker.__init_subclass__() takes no keyword arguments

Verified by installing the package against faststream==0.7.5. SubscriberSpec/PublisherSpec
also lack address there. The floor moves 0.7.1 -> 0.7.6 in both the runtime dependency and the
dev group. Nothing in CI resolves at the lower bound, which is why the old floor went stale without
anyone noticing; worth a UV_RESOLUTION=lowest-direct job, but that is not in this PR.

Compatibility

Runtime behaviour is unchanged for anyone already on faststream 0.7.6. Three things are worth
calling out:

• Installs on faststream < 0.7.6 are now refused by the resolver rather than failing at import.
  That is the point of the floor bump, and it is a real narrowing for anyone pinning 0.7.1-0.7.5,
  though such an install is already broken by 0.7.6's own API removals.
• dependencies= no longer type-checks for a non-Sequence iterable (a generator, a set).
  Runtime is unaffected, there is no isinstance check; only type checkers will complain. Matches
  what upstream did to the same kwarg.
• The generated AsyncAPI document gains an address on every channel. Additive: 0.7.5's
  SubscriberSpec/PublisherSpec had no such field. Real queue names everywhere except a
  multi-queue subscriber, which gets the comma-joined string until AsyncAPI: emit one channel per queue, not one comma-joined channel per subscriber #181 lands.

__aiter__ switching from async def to def is not a break in practice: it raised
NotImplementedError before and raises it now, just from the call rather than the await, so
await sub.__aiter__() still behaves identically.

Verification

Run locally against Postgres 17 via compose, so test_integration.py was not skipped.

• eof-fixer --check, ruff format --check, ruff check --no-fix, ty check all clean
• 623 passed, coverage 100.00% (gate holds)
• mkdocs build --strict clean
• python -m benchmarks check: gate OK

[github-actions]

Benchmark gate

✅ gate passed

scenario	msg/s	delete/msg	WALrec/msg	WALB/msg	fpi	upd	del	dead_tup
consumer/w1/b10	1108	1.000	6.81	909	242	5000	5000	10000
consumer/w1/b100	896	1.000	6.88	917	243	5000	5000	10000
consumer/w2/b10	1383	1.000	6.79	958	242	5000	5000	10000
consumer/w2/b100	1678	1.000	6.73	955	243	5000	5000	10000
consumer/w4/b10	1303	1.000	6.74	962	261	5000	5000	10000
consumer/w4/b100	2011	1.000	6.86	1033	244	5000	5000	10000
consumer/w1/b100/tfbs100	8260	0.010	6.07	1116	243	5000	5000	10000
producer/w1/b100	2942	0.000	3.04	584	0	0	0	0

Gated (fails the build): delete_calls + tuple counters (upd/del/ins) + the producer's insert_calls, exact; select_calls within +2; wal_records within a 10% band. msg/s, WAL bytes and total calls are informational (timing/FPI noise).